We work in tight interdisciplinary teams, each focused on a single goal at a time.

We try to expand our modeling capabilities incrementally with each release: every model attempts something new while building on accumulated knowledge from previous systems. This philosophy was partly inspired by discussions around how the Apollo program expanded capabilities incrementally rather than attempting the full problem at once.

All groups operate as interdisciplinary units working toward a single focused objective. We try to minimize fragmentation of attention and maximize iteration speed. Collaborations allow us to collectively explore a much broader range of ideas and research directions than any single team could pursue alone. We are therefore excited to support ambitious computational biology “moonshots” from the community, especially when aligned with our current Open Problems.

Kristof Szalay, CTO

Drug Combination Group

Lead: Robert Sipos

Members: Dana Zemel Kristof Galambos Murat Cem Kose

We optimize models for drug combinations to predict possible synergies in a vast search space. Predicting synergy is a much more difficult task compared to predicting responses for mono treatments due to the dramatically worse signal-to-noise ratio. This also makes data harmonization exceptionally important.

Another focus of our group is modeling advanced modalities, like antibody-drug conjugates (ADCs) or degraders like PROTACs. Utilized in Turbine's current ADC product, we developed solutions to match antibodies with the best payload for a given indication.

In vivo Biology Group

Lead: Imre Gaspar

Members: Valér Kaszas Balint Pfliegel Lucie Studena

In vivo data is as scarce as it is expensive to generate. The In vivo Biology Group focuses on the transfer learning problem of using in vitro perturbation data (CRISPR, small molecule, etc., mainly from HT screens) to build in vivo predictive ML/AI models. Our approach is to harmonize different biosample domains through decomposition and in-model resynthesis that allows the model to learn from sample similarities while still being capable of distinguishing between the in vivo and in vitro domains.

Multimodal Learning Group

Lead: Laszlo Mero

Members: Csaba Gor Milan Sztilkovics Zsolt Gyüre Alex Sliz-Nagy Peter Frendl

In vitro data is diverse with multiple perturbation modalities and many possible endpoint measurements. We focus on enabling models to learn the underlying biology with high fidelity by combining datasets from different modalities. Our current project is creating a model for target discovery that combines CRISPR screens, PerturbSeq screens and drug modifier CRISPR data.

Beyond Oncology Group

Lead: Bence Szalai

Members: Bence Czako Gema Sanz Gerold Csendes

The Beyond Oncology Group specializes in applying virtual assay capabilities to new therapeutic areas beyond oncology, with a particular focus on immune system.

Our work is grounded in the idea that (perturbation) RNA-seq - bulk, single-cell, and spatial - provides a robust, scalable representation of cellular state, even outside traditional oncology settings. This approach is reflected in our participation in the Virtual Cell Challenge, where our "Mean Predictors" team achieved top-tier performance.

Our best work is only possible within the broader ecosystem of Turbine.

We'd like to acknowledge the work of related Turbine teams without whom our work would not be possible:

• Lab for generating valuable data that cannot be found anywhere else
• Engineering for providing us with the data and compute infrastructure
• Product and partnership teams for keeping our work grounded in real-life applications
• Marketing for helping us communicate our ideas
• and support teams for helping us keep the teams healthy.

Thank you for providing the psychological and institutional safety that allows us to move quickly and speak honestly about our limitations and failures.